Livealgos

<h2 align="center">The World's First Live Open-Source Trading Algorithm<BR> [Note: This Site Is Not Functional Yet]</h2> <p align="center"> <a href="https://livealgos.com"><img src="original_image.jpg" alt="LiveAlgos.com Logo" Height="200" width="350"></a> </P> <p align="center"> Use our money to test your automated stock/FX/crypto trading strategies. All changes users make to our Python GitHub code are added to the repo, and then reflected in the live trading account that goes with it. You can also use that same code to trade with your own money. We have spent over 10 years developing automated trading strategies, and are open sourcing all of our programming (meaning it is free for anybody to use) in the hopes that users will help improve it. Here is some of what we have learned so far from doing automated trading: </p> <ul> <li>Backtests are great to use for developing strategies but are somewhat meaningless in predicting actual profits. The only way to know how well your strategy works is to trade it live. </li> <li>There are all sorts of issues that come up in live trading which are not reflected accurately in backtesting (or paper trading). For example, even if you do a good job estimating the brokerage fees and the bid/ask spread, they will almost always end up being worse in real trading. Limit orders can help with this slippage in live trading but are almost impossible to backtest because they would almost never get filled on the price the backtester shows.</li> <li>It is very hard to go from backtesting to live trading. Most backtesting systems do not support live trading, at least not without additional programming. </li> <li>Very few backtesting and trading programs support machine learning (using artificial intelligence to predict the price). They only allow you to create a strategy using the standard list of indicators (moving average, Bollinger bands, RSI, etc.). </li> <li>The best way to combine multiple strategies is to use machine learning (ML). ML automatically determines which strategies are best to use at what times, and can be easily retrained with new data. Otherwise, strategies that backtested well may work for a little while and then become old and stop working.</li> <li>High frequency trading (HFT) is only worth pursuing if you are trading millions of dollars and are willing to spend thousands of dollars a month on quant technologies. This is not something we are interested in doing.</li> </ul> </div>

There are 2 main ways to improve our existing ML trading algorithm:

<ol> <li>Increase the accuracy by trying new machine learning methods. For example, there are always new types of neural networks that could be tested, or new time series prediction libraries that could be used. That similar to how they do it in the contest at <A HREF="http://numer.ai" target="new">Numerai</a>, but the problem with Numerai is that all of the data is anonymous and encrypted. You have no idea what financial instruments and indicators/features you are working with. So if you come up with something good, there is no easy way to apply it to real trading for yourself. </li> <li>Add new features to help the ML model learn better. Features can be new technical indicators or a complete strategy that gives buy and sell signals. The ML combines these new indicators and strategies with all of the existing ones, to improve on the model it already has. This is similar to the contests at <A HREF="https://www.kaggle.com/competitions?sortBy=relevance&group=general&search=stock&page=1&pageSize=20" target="new">Kaggle.com</a>, but Kaggle rarely deals with stocks, and when they do it, it is still hard to apply the results to real trading. </li> </ol> <h2>What We Have Working So Far</h2> <br> <UL> <LI>Backtester with stock, FX, and crypto data.</LI> <LI>100+ indicators from TA-Lib (ta-lib.org) plus some we created ourselves, added as features. Also various time series features such as ARIMA, SARIMA, ARIMAX, and SARIMAX.</LI> <LI>Optimization of the 11,000+ indicator paramater combinations, using feature reduction to find which ones work best.</LI> <LI>Over 60 classification and regression algorithms and neural networks using our custom made AutoML program.</LI> <LI>ML model parameter optimization using Skopt, genetic algorithm, or exhaustive search. <LI>Voting ensembles and stacking of algorithms to acheive higher accuracy.</LI> <LI>Tests using evolutionary algorithms (such as NEAT) and genetic programming (such as gplearn).</LI> <LI>Tests using reinforcement learning algorithms such as policy gradient, Q-learning, evolution strategy, actor-critic, curiosity Q-learning, and neuroevolution with novelty search.</LI> <LI>Over a dozen dimensionality reduction techniques for selecting the best features, such as PCA, RandomTreesEmbedding, LDA, SelectKBest, SelectFromModel, SymbolicTransformer, GeneticSelectionCV, SNERBFSampler, RFECV, FastICA, Isomap, SpectralEmbedding, FeatureAgglomeration, and LocallyLinearEmbedding.</LI> <LI>Coming Soon - Over 1000 new features from various MetaTrader indicators and strategies. These can be used on anything, not just FX.</LI> <LI>Coming Soon - Live trading. We have the ability to trade live right now, but we want to make the algorithm more profitable before we do that.</LI> </UL>

Below are the accuracy results (classification) on unseen test data using 4 years of hourly EURUSD data, with a 75%/25% train/test split:<BR> [coming soon]<BR> And here's the backtest results (exit trade at the end of each 1 hour bar, no commissions/slippage) on the unseen test set, using XGBoost:<BR>

<p align="left"> [coming soon] </P> <BR> <B>Program Requirements:</B><BR> Linux Server (we use Ubuntu 16.04)<BR> Python 3.6<BR> There are many other smaller packages that you will need also, see the requirements.txt file.<BR> <BR> Our Programs Consists of 4 Main Python Notebooks:<BR> <BR> <B>Data Handling and Feature Generation</B> - datamodel_dev.ipynb - Loads the raw price data, preprocesses it (drop NANs, normalization, etc.), balances the dataset (equal number of ups and downs), and adds new features (such as moving average, Bollinger Bands, time lags etc.).<BR> <BR> <B>Feature Selection</B> - Uses dimensionality reduction (such as PCA) to find the most useful of the 11000+ features from the data file. This helps the ML learn better by letting it focus on what is really important, instead of distracting it with useless noise. And, it makes everything run much faster to use less data. We run over 20 different feature selection programs and then use XGBoost on the datasets they generate, to see which has the highest accuracy. This narrows it down from using over 11,000 features to only the top 100.<BR> <BR> <B>Algorithm Selection</B> - Test 60 different ML algorithms on the data to see which one gives the highest accuracy.<BR> <BR> <B>HPO and Ensembles</B> - Optimizes the parameters of the ML model (such as XGBoost) to increase accuracy. It then runs it 100 times to create an ensemble model, giving it more stability.<BR> <BR> Note: The RL programming was done by Peter Chervenski (https://www.facebook.com/l.s.d.records), with assistance from Eric Borgos. The "Roadmap For The Future" part below is by Eric Borgos. <h2>Roadmap For The Future</h2> <br> Suggested improvements for users to try:<BR><BR> <OL> <LI><B>Use these programs to get new features:</B><BR> TSFresh: <a href='https://github.com/blue-yonder/tsfresh' target = "new">https://github.com/blue-yonder/tsfresh</a><BR> Cesium: <a href='https://github.com/cesium-ml/cesium' target = "new">https://github.com/cesium-ml/cesium</a><BR> PyAF: <a href='https://github.com/antoinecarme/pyaf' target = "new">https://github.com/antoinecarme/pyaf</a><BR> TSLearn - <a href='https://github.com/rtavenar/tslearn' target = "new">https://github.com/rtavenar/tslearn</a><BR> pyts: <a href='https://github.com/johannfaouzi/pyts' target = "new">https://github.com/johannfaouzi/pyts</a><BR> Time Series Feature Extraction Library: <a href='https://github.com/fraunhoferportugal/tsfel' target="new">https://github.com/fraunhoferportugal/tsfel</a><BR> Khiva: <a href='https://github.com/shapelets/khiva' target = "new">https://github.com/shapelets/khiva</a> (C++) or <a href='https://github.com/shapelets/khiva-python' target = "new">https://github.com/shapelets/khiva-python</a> (python bindings)<BR> PyFTS: <a href='https://github.com/PYFTS/pyFTS' target = "new">https://github.com/PYFTS/pyFTS</a><BR> Genetic Discovert of Shapelets: <a href='https://github.com/IBCNServices/GENDIS' target = "new">https://github.com/IBCNServices/GENDIS</a><BR> PyFlux: <a href='https://github.com/RJT1990/pyflux' target = "new">https://github.com/RJT1990/pyflux</a><BR> Deep Learning for Time Series Classification: <a href='https://github.com/cauchyturing/UCR_Time_Series_Classification_Deep_Learning_Baseline' target = "new">https://github.com/cauchyturing/UCR_Time_Series_Classification_Deep_Learning_Baseline</a><BR> PMDARIMA: <a href='https://github.com/tgsmith61591/pmdarima' target = "new">https://github.com/tgsmith61591/pmdarima</a><BR> Stumpy: <a href='https://github.com/TDAmeritrade/stumpy' target = "new">https://github.com/TDAmeritrade/stumpy</a><BR> Feets: <a href='https://github.com/carpyncho/feets' target = "new">https://github.com/carpyncho/feets</a><BR> Thalesians' Time Series Analysis (TSA) - <a href='https://github.com/thalesians/tsa' target = "new">https://github.com/thalesians/tsa</a> (notebooks at <a href='https://github.com/thalesians/tsa/tree/master/src/jupyter/python' target = "new">https://github.com/thalesians/tsa/tree/master/src/jupyter/python</a>) - Made specifically for stock time series analysis.<BR> SLearn: <a href='https://github.com/mzoll/slearn' target="new">https://github.com/mzoll/slearn</a><BR> TimeNet: <a href='https://github.com/kirarenctaon/timenet' target="new">https://github.com/kirarenctaon/timenet</a><BR>